PROJECT SUMMARY/ABSTRACT Tools to determine and analyze the structures of molecular machines in motion Single particle cryo-electron microscopy (cryo-EM) has transformed our ability to rapidly determine high resolution structures of static, structurally homogeneous macromolecular complexes. However, we have not realized cryo-EM’s potential to uncover the full ensemble of heterogeneous structures these molecules adopt as they function. The overall objective of this work is to develop novel cryo-EM image processing tools to: 1) determine the complete ensemble of structural states adopted by imaged complexes; 2) quantify the relative abundance of these states; 3) monitor how the distribution of these states changes as the machine functions; and 4) use this information to understand the molecular mechanism of how these machines assemble and function. This objective is important as visualizing structural ensembles can be vital in developing and testing hypotheses for how these machines function, and in developing therapeutics to modulate their activity. Here, we specifically aim to develop two tools to facilitate achieving these overall objectives. First, we will generate ‘benchmark’ datasets that will be distributed to the methods development community to aid in building and quantitatively assessing of the fidelity of different approaches to reconstruct 3D density maps from single particle cryo-EM data. These benchmark datasets will include macromolecular complexes bearing elements of structural heterogeneity we have specifically designed for this purpose, and that we have biochemically assembled and imaged. Additionally, it will design, implement, and validate a machine learning-based computational tool that more realistically simulates the imaging process than existent software, thereby enabling users to rapidly construct custom synthetic benchmark datasets to test specific aspects of their own algorithms. Recently, as a proof-of-concept, we published the first method using deep neural networks to perform 3D reconstruction from single particle data, and this approach was particularly efficacious is revealing heterogeneous structures. Thus, our second aim is to develop this approach into a complete software package enabling users to readily reconstruct hundreds-to-thousands of density maps from a single dataset; to implement tools to focus the analysis on specific structural regions; and to deploy methods guiding the interpretation of the density maps and the construction of ensembles of associated atomic models. This work in innovative in its objective to analyze heterogeneous structural ensembles as opposed to static structures at high resolution; in our approach to model model conformational changes as originating from a continuous distribution of structures as opposed to isolated, discrete states; and in our application of deep learning methods to both the generation of benchmark datasets and in the reconstruction process itself. As a proof-of-c...